Elsevier

NeuroImage

Volume 39, Issue 1, 1 January 2008, Pages 359-368
NeuroImage

Simulation-based evaluation of OSEM iterative reconstruction methods in dynamic brain PET studies

https://doi.org/10.1016/j.neuroimage.2007.07.038Get rights and content

Abstract

The reconstruction of dynamic PET data is usually performed using filtered backprojection algorithms (FBP). This method is fast, robust, linear and yields reliable quantitative results. However, the use of FBP for low count data, such as dynamic PET data, generally results in poor visual image quality, exhibiting high noise, disturbing streak artifacts and low contrast. These signal-to-noise ratio and contrast in the reconstructed images may alter the quantification of physiological indexes, such as the regional Binding Potential (BP) obtained from kinetic modeling. Iterative reconstruction methods are often presented as viable alternatives to FBP reconstruction. In this study, we investigated the characteristics of the UW-OSEM and the ANW-OSEM iterative reconstruction methods in the context of ligand–receptor PET studies with low counts. The assessment was conducted using replicates of simulated [18F]MPPF acquisitions. The quantitative accuracy obtained with the iterative and analytical methods was compared. The results show that analytical methods are more robust to the low count data than iterative methods, and therefore enable a better estimate of the regional activity values and binding potential. The positivity constraint in MLEM-based algorithms leads to overestimations of the activity in regions with low activity concentration, typically the cerebellum. This overestimation results in significant bias in BP estimates with iterative reconstruction methods. The bias is confirmed from the reconstruction of real PET data.

Introduction

The radiotracer 2′-methoxyphenyl-(N-2′-pyridinyl)-p-18F-fluoro-benzamidoethylpiperazine ([18F]MPPF) is a specific serotonin 5-HT1A antagonist PET tracer recently characterized, modeled, and used for clinical research to explore abnormalities in the serotoninergic system (Merlet et al., 2004a, Merlet et al., 2004b). The in vivo exploration of the 5-HT1A receptors with [18F]MPPF PET imaging is of great interest as those receptors are involved in numerous neurological and psychiatric disorders. Similar to other ligand–receptor PET studies, the acquisition protocol consists in the collection of coincidence events into multiple short time frames over a total scanning duration ranging from 60 to 90 min. Whereas iterative reconstruction methods are available with commercial scanners, Filtered Back Projection (FBP3D) (Kinahan and Rogers, 1989) is usually preferred for the reconstruction of dynamic brain PET scans. Indeed, this reconstruction method is fast, robust, linear and yields reliable quantitative results. Iterative reconstruction methods are often based on the Maximum-Likelihood Expectation–Maximization (MLEM) algorithm (Shepp and Vardi, 1982). Accelerated MLEM-based reconstruction algorithms such as Unweighted Ordered Subsets Expectation–Maximization (UW-OSEM) (Hudson and Larkin, 1994) and more specifically Attenuation-Weighted OSEM (AW-OSEM), Attenuation Normalization-Weighted OSEM (ANW-OSEM) (Michel et al., 1998) and Ordinary Poisson OSEM (OP-OSEM) (Yavuz and Fessler, 1996) include some data correction within the iteration process and better account for the nature of the noise. They are often presented as viable alternatives to FBP reconstruction.

Many comparison studies showed that iterative reconstruction outperforms FBP in terms of image quality, signal-to-noise ratio, resolution and contrast (Bouchareb et al., 2005, Riddell et al., 2001, Gutman et al., 2003, Wang et al., 1998, Boellaard et al., 2001), and improves lesion detection (Lartizien et al., 2003). They highlighted that the characteristics of the reconstructed images are bound to the chosen number of iterations and to the source distribution (Gutman et al., 2003, Wang et al., 1998). Consequently, for a specific PET protocol, the number of iterations must be carefully selected so as to achieve reliable quantitative results while limiting the noise amplification. Most comparison studies were conducted using [18F]FDG static scans containing relatively high numbers of detected events as in the context of tumor detection. Consequently, little is known about the characteristics of iterative reconstruction techniques of PET data in low count rate situations.

In this study, we evaluate the characteristics of the UW-OSEM and ANW-OSEM reconstruction methods operating on 3D data and of UW-OSEM and AW-OSEM on 2D data after Fourier rebinning (FORE) (Defrise et al., 1997), using simulated ligand–receptor [18F]MPPF PET studies. In this application, the accuracy and variability of the activity levels and kinetic parameter estimates are of a higher concern than the visual quality of the reconstructed images. We compared the performance of the two iterative methods with performance obtained with analytical methods, i.e. FORE + FBP2D, FBP3D and the Direct Inversion Fourier Transform method (FORE + DIFT) (Matej and Bajla, 1990). The motivation for conducting this study was to answer the practical question, “Is OSEM reconstruction a viable alternative to FBP reconstruction in quantitative PET studies in the case of low count data?” It is somewhat similar to few previously published studies such as Michel et al. (1999) for benzodiazepine brain PET studies, Morimoto et al. (2006) for [11C]raclopride and [11C]DASB, Bélanger et al. (2004) for [11C]WAY, and Koch et al. (2005) for dopamine SPECT studies. In our context, they provided inconclusive results. The originality of our methodology compared to previously published studies stems from the use of multiple replicates of realistic simulated [18F]MPPF PET scans to accurately assess the impact of noise on the performance on the reconstructed methods. This choice was motivated by the fact that there only exists approximate local variance estimators for nonlinear reconstruction algorithms (Barrett et al., 1994, Fessler, 1996, Kadrmas et al., 1999, Qi and Leahy, 2000, Soares et al., 2000, Wang and Gindi, 1997, Wilson et al., 1994, Buvat, 2002), and that experimentally acquired replicates using phantoms or gated acquisitions as proposed by Riddell et al. (2001) do not provide the required flexibility to model very realistic dynamic acquisitions.

Section snippets

Material and methods

The strategy we used to assess the impact of low count data on the accuracy of the different reconstruction methods consists of comparing the quantitative measurements obtained from the reconstruction of the scans with a statistic varying from low (normal) to high (11 times the normal level). Replications of the same PET acquisition protocol (transmission + dynamic emission) were performed.

Optimal number of iterations and subsets

The impact of the number of subsets (4 and 16) was estimated by comparing the kinetic parameters BP calculated for the 3 regions using the PET volumes obtained with the different reconstruction methods ({ANW, UW}-OSEM  {2D,3D}  {16,4}subsets). Overall, the results show that beyond 80 MLEM equivalent iterations (iterations × subsets), the number of subsets has no impact on the determination of the kinetic parameter values. Indeed, after 80 iterations, the observed relative differences between BP

Discussion

In this study, using multiple data replicates, we investigated the performances of various iterative reconstruction schemes in 3D and in 2D after FORE, in case of low count data. Different conclusions can be drawn from the presented results.

Conclusion

We investigated the performance of the UW-OSEM and ANW-OSEM iterative reconstruction methods of data containing low statistics using multiple simulated replicates. The results showed that the studied iterative reconstruction methods are biased at low statistics, especially in the lower part of the image dynamic range and for cold regions. The bias comes from the lack of detected events rendering the data prone to negative values after corrections along with the zero-thresholding in the sinogram

Acknowledgments

The authors would like to thank the following people for discussions and assistance in this study:

  • Fabrice Bellet and Christophe Pera (CREATIS, Lyon, France) for providing us the computational resources.

  • Carole Lartizien (CREATIS, Lyon, France) and Andrew Reader (School of Chemical Engineering and Analytical Science, University of Manchester, UK) for helpful discussions.

References (41)

  • I. Buvat

    A non-parametric bootstrap approach for analysing the statistical properties of SPECT and PET images

    Phys. Med. Biol.

    (2002)
  • C. Byrne

    Iterative algorithms for deblurring and deconvolution with constraints

    Inverse Problems

    (1998)
  • J.-C.K. Cheng et al.

    A scatter-corrected list-mode reconstruction and a practical scatter/random approximation technique for dynamic PET imaging

    Phys. Med. Biol.

    (2007)
  • M. Defrise et al.

    Exact and approximate rebinning algorithms for 3-D PET data

    IEEE Trans. Med. Imag.

    (1997)
  • K. Erlandsson et al.

    Low-statistics reconstruction with AB-EMML

    IEEE Nucl. Sci. Symp. Conf. Rec.

    (2000)
  • J.A. Fessler

    Mean and variance of implicitly defined biased estimators (such as penalized maximum likelihood): applications to tomography

    IEEE Trans. Image Process.

    (1996)
  • F. Gutman et al.

    Optimisation of the OS-EM algorithm and comparison with FBP for image reconstruction on a dual-head camera: a phantom and a clinical 18F-FDG study

    Eur. J. Nucl. Med. Mol. Imaging

    (2003)
  • H.M. Hudson et al.

    Accelerated image reconstruction using ordered subsets of projection data

    IEEE Trans. Med. Imag.

    (1994)
  • D.J. Kadrmas et al.

    Analytical propagation of errors in dynamic SPECT: estimators, degrading factors, bias and noise

    Phys. Med. Biol.

    (1999)
  • P.E. Kinahan et al.

    Analytic 3D image reconstruction using all detected events

    IEEE Trans. Nucl. Sci.

    (1989)
  • Cited by (0)

    View full text